SmoothConv

SmoothConv is a high-quality Chinese multi-channel conversational speech dataset with expert human annotations, developed by ASLP@NPU and QualiaLabs as part of the SmoothConv–DuplexConv corpus family.

Companion dataset: DuplexConv on HuggingFace (2,000 hours, LLM-assisted annotation). SmoothConv and DuplexConv are constructed from the same underlying conversational sources. SmoothConv provides high-fidelity human annotations for benchmarking and supervised training; DuplexConv offers large-scale annotations for Speech LLM pre-training and data-driven modeling.

Dataset Overview

SmoothConv contains 100 hours of naturally occurring multi-party Chinese conversations recorded in multi-channel environments across Tutoring and Social Chat scenarios. Unlike corpora dominated by read speech or scripted interactions, it captures realistic conversational dynamics, including overlapping speech, backchannels, interruptions, pauses, and turn transitions.

The dataset is manually annotated by trained experts and provides fine-grained conversational labels, making it suitable for turn-taking modeling, overlap and interruption detection, full-duplex spoken dialogue systems, conversational speech understanding, and Speech LLM research.

Metric	Value
Total Duration	100.53 hours
Audio Files	2,503
Mean Duration	144.59 sec
Duration Range	60.0 – 634.7 sec
Language	Chinese (zh)
Domains	Tutoring, Social Chat
Annotation	Expert human annotation

Domains & Directory Layout

After download, each conversation is stored under a top-level folder whose name indicates the scenario. Match the folder prefix to the domain:

Scenario	Folder prefix	Example
Tutoring	starts with edu or Edu	Edu_20240101_001/
Social Chat	starts with none_Edu	none_Edu_20240101_001/

Within each folder you will find paired multi-channel audio (.wav) and annotation (.json) files. The same naming convention applies to both SmoothConv and DuplexConv.

Dataset Statistics

Turn-taking labels include complete, incomplete, backchannel, and wait.

Supported Tasks

• Turn-taking modeling
• Overlap and interruption detection
• Full-duplex spoken dialogue systems
• Conversational speech understanding
• Speech Language Models (Speech LLMs)

Annotation Format

Each audio file is paired with a JSON annotation file. The top-level object contains an instances list; each element describes one annotated segment.

Top-level structure

Field	Type	Description
instances	list	List of annotated segments in the conversation

instances[i] — per-segment annotation

Field	Type	Description
id	str	Unique segment identifier (UUID)
channelIndex	int	Audio channel index (0-based)
start	float	Segment start time in seconds
end	float	Segment end time in seconds
text	str	Human-annotated transcript; inline tags mark events (e.g. <pause>, <噪声>, <unclear>)
attributes	dict	Speaker, turn, paralinguistic, and scene attributes (see below)

instances[i].attributes

Field	Type	Description
speaker	str	Speaker ID (e.g. A1, B1); unknown if unidentifiable
turn	str	Turn-taking state: complete, incomplete, backchannel, wait
other_turn	list (optional)	Co-occurring interaction cues, e.g. pause, unknown turn
gender	str	Speaker gender
age	str	Speaker age group (e.g. adult, child)
emotion	str	Emotion label for the segment
speech event	list	Paralinguistic / non-speech events (e.g. nonespeech event, echo, shouting)
这段数据是在什么环境	str	Scene / environment description

Example segment

{
  "id": "0d0687e7-b2e5-4b91-834b-f3e8988e7a4a",
  "channelIndex": 0,
  "start": 0.702,
  "end": 5.146,
  "attributes": {
    "speaker": "A1",
    "turn": "complete",
    "gender": "male",
    "age": "adult",
    "emotion": "neutral",
    "speech event": ["nonespeech event"],
    "这段数据是在什么环境": "unknown"
  },
  "text": "春风花草香迟日江山丽日出江花红胜火"
}

Usage

import json
from datasets import load_dataset

ds = load_dataset("qualialabsAI/SmoothConv")

# Load annotation for a sample
with open("path/to/annotation.json", "r", encoding="utf-8") as f:
    anno = json.load(f)

for seg in anno["instances"]:
    print(seg["channelIndex"], seg["start"], seg["end"], seg["text"])

Ethics Statement

• Informed consent. Conversations were recorded with the knowledge and consent of participants. Personal identifiers have been removed or anonymized prior to release.
• Privacy protection. For academic and research use only. Do not attempt to re-identify speakers or reconstruct private information.
• Intended use. Research on spoken dialogue, turn-taking, and speech understanding—not for unauthorized surveillance, impersonation, or deceptive content generation.
• Limitations & bias. Human annotations may contain errors; account for domain and demographic bias in experiments.
• Responsible use. Report suspected misuse to jimz@qualialabs.ai.

License

CC BY-NC 4.0

Citation

@article{wang2026duoconv,
  title   = {DuoConv: Large-Scale Chinese Full-Duplex Speech Datasets for Conversational AI},
  author  = {Chengyou Wang and Chunjiang He and Zhou Zhu and Lei Xie},
  journal = {arXiv preprint arXiv:0000.00000},
  year    = {2026},
  note    = {Placeholder; paper forthcoming}
}

Contact

jimz@qualialabs.ai